Topology-Aware Safe Deep Reinforcement Learning for Resilient Energy Management of Uncertain Active Distribution Networks

The increasing penetration of photovoltaic generation, battery energy storage systems, and electric vehicles is transforming conventional distribution networks into active and increasingly uncertain cyber-physical systems. The resulting variability in generation and demand, together with changing network operating conditions, creates significant challenges for real-time energy management. Deep reinforcement learning (DRL) has emerged as a promising approach for controlling active distribution networks because a trained policy can generate control decisions without repeatedly solving a computationally intensive optimization problem. However, conventional DRL methods generally represent the electrical network using fixed-dimensional vectors and may therefore fail to adequately exploit the topological relationships between buses and branches. In addition, an unconstrained DRL policy may generate actions that violate electrical and energy-storage operating limits. This paper proposes a topology-aware safe deep reinforcement learning framework for resilient energy management of uncertain active distribution networks. The proposed framework integrates a Graph Neural Network (GNN) with the Soft Actor-Critic (SAC) algorithm to learn topology-aware representations of the electrical system. A safety projection mechanism is further introduced to transform potentially infeasible control actions into feasible actions before they are applied to the network. To improve robustness against uncertainty, domain randomization is employed during the training process. The proposed approach is evaluated using a physics-based IEEE ۳۳-bus active distribution network under normal operating conditions, renewable-generation uncertainty, load uncertainty, and topology disturbances. The performance is compared with conventional SAC and vector-based reinforcement-learning approaches using operating cost, voltage violations, branch loading, battery utilization, computational efficiency, and generalization to unseen network conditions. The experimental results will be reported after implementation and simulation, without introducing fabricated numerical results.

کلیدواژه ها:

Active distribution network ، Graph Neural Network ، Safe reinforcement learning ، Soft Actor-Critic ، Energy management ، Battery energy storage ، Electric vehicles ، Resilience ، Uncertainty ، Smart grid